Seite - 213 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments
Bild der Seite - 213 -
Text der Seite - 213 -
tem monitoring such databases and as insurance, copies of almost each data pinned to
a system file are a necessity for the correct functioning of a database. Erasure, in this
case, appears as either going around those points where the intended data for erasure is
stored and a query is made upon, and deleting something from logs or backups is almost
impossible (also not recommended). Moreover, there is no real erasure, data is only ear-
marked as removed, without actually being erased. In practice, in these cases, the query
will no longer look for that specific data in the search index because of this earmark. In
AI systems where each data is evaluated and fed back into the Machine Learning (ML)
process as training data, we think that it is not possible to pull a single data from such a
structure. Authors come to such a conclusion that the GDPR may affect the development
of ML techniques in Europe because of legal obligations misfit to technological realities.
By assuming the existence of AI systems ability to forget data, the other side of
the coin makes us assume that they may have an ability to remember. Carlini et. al.[30]
admits that data memorization starts at the early phase of training the Neural Networks
and it is hardly an unavoidable common issue even if it is unintended (meaning where
the data is not useful for learning task and neither for the accuracy of the model). The
researchers prove that there are solutions for avoiding the unintended memorization, al-
though they reach such conclusion by testing white-box algorithms at a small cost. Ap-
plying differential privacy methods may make the algorithm learn slower than standard
and may cause utility loss. We are not sure whether such losses would also affect the
costs for AI development, but if so, these safeguards should be implemented in the de-
velopment phase anyhow. The GDPR is quite well prepared in this sense, since Data
Protection by Design and by Default (Art. 25) was inserted in the legal text. However, as
the European Data Protection Supervisor also refers to it in its preliminary opinion [31],
the wording of this article does not include the developers, only the data controllers who
may not necessarily get involved with the early phase of the AI systems development.
Moreover, general recommendations made for specific technologies like AI may not al-
ways cover all possible scenarios, so more specific recommendations and explanations
should be delivered by the EDPS regarding implementation of the GDPRs Article 25 on
AI technologies.
3. Conclusion
In this paper, we focused on the practicability of the consent rule in connection with
many rights and principles regarding data protection, but especially with the principle of
transparency and RTBF. Data controllers whose obligation is to deliver transparent infor-
mation to the data subjects may find themselves in a difficult position to present univer-
sally understandable (intelligible) information regarding their AI systems. The GDPR,
however, does not order data controllers to verify the understandability of the informa-
tion they present to data subjects. Consequently, some data controllers may use the trans-
parency rule to trick data subjects, some of them may undertake unrealistic commit-
ments not matching their actual capabilities and practices. Consent is also interrelated
with RTBF, as the lack of consent obliges data controllers to make the AI system forget
upon the request of data subject. However, as we illustrated, AI systems may not be able
to forget in a way as the GDPR intends RTBF to function. Technical and practical prob-
lems implementing the GDPR on AI systems together with legal uncertainties may con-
G.GultekinVarkonyi /Operability of theGDPR’sConsent Rule in Intelligent Systems 213
Intelligent Environments 2019
Workshop Proceedings of the 15th International Conference on Intelligent Environments
- Titel
- Intelligent Environments 2019
- Untertitel
- Workshop Proceedings of the 15th International Conference on Intelligent Environments
- Autoren
- Andrés Muñoz
- Sofia Ouhbi
- Wolfgang Minker
- Loubna Echabbi
- Miguel Navarro-Cía
- Verlag
- IOS Press BV
- Datum
- 2019
- Sprache
- deutsch
- Lizenz
- CC BY-NC 4.0
- ISBN
- 978-1-61499-983-6
- Abmessungen
- 16.0 x 24.0 cm
- Seiten
- 416
- Kategorie
- Tagungsbände